Detectron: Using tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml to train VOC2007

[zhanglonghao1992]

I download PASCAL VOC Annotations in COCO Format through "Detectron/lib/datasets/data/README.md".
Then I train VOC2007 by running "Detectron/configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml" .I have changed NUM_CLASSES to 21, training DATASETS to ('voc_2007_train'), testing DATASETS to('voc_2007_test'),and I set SCALES to (600,),MAX_SIZE to 1000. After 100000 iterations,I got a very low mAP,about 0.69.I dont know whats wrong.
Someone help~

[YoungMagic]

Same thing happened training custom dataset. I'm working on this.

[zhanglonghao1992]

@YoungMagic Cant find how to adjust those parameter settings in this yaml file to get a high mAP using only one GPU. With caffe1,I can easily get a good mAP using just one GPU.

[JackieZhangdx]

Similar problem. e2e_mask_rcnn_R-50-FPN.yaml, training DATASETS ('voc_2012_train'), testing DATASETS to('voc_2012_val'), 8GPU with 45000 iter LR start from 0.02
mAP 0.65 for detection

[YoungMagic]

I found that e2e_faster_rcnn_R-50-C4 has a far better result than R-50-FPN……

[zhanglonghao1992]

@JackieZhangdx Have you solved that？

[JackieZhangdx]

@zhanglonghao1992 you may try larger training datasets such as coco or VOC07++12 to train, which boosts the performance evidently. Or other data augmentation methods as the Detectron only employ flipping by default( because it tackles coco dataset whose training set is much larger.)

[zhanglonghao1992]

@JackieZhangdx 3Q，I am trying coco2014

[zhonhel]

@zhanglonghao1992
Hi, the result of your post, the 'voc_2007_train' means you only used the voc 2007 train set to train, and haven't added the voc 2007 val set ?
I want to reproduce your result.
Thank you very much!

[jy1023408440]

I solved it,detail https://blog.csdn.net/jy1023408440/article/details/89474759